📦 calmrocks/ai-engineer-notebooks

Build Real AI Systems for Free

Build working AI systems from raw code using the AI Engineer Notebooks, a free collection of Colab projects that skips the heavy frameworks to teach you the actual mechanics of modern AI. This repository is a hands-on curriculum that runs entirely on a fast, no-cost API, letting you construct everything from basic prompts to complex agents and retrieval systems without spending a dime. The real magic is in the final section, where you don't just read about production issues, but actually debug a live support assistant as it fails in the real world. It’s the perfect way to prove you can ship real software, not just demo code.

📰 https://news.ycombinator.com/item?id=49471714

🆔 @hackernewsgithubprojects

📦 real-stanford/gated-memory-policy

Gated Memory Policy: Robots That Remember

A robot arm can now watch a human work and then execute the task in its own kitchen without a single line of code retraining. This Stanford project, Gated Memory Policy, teaches machines to store what they see in a flexible internal memory, allowing them to adapt to new situations on the fly. Instead of treating every movement as a blank slate, the system remembers past actions and adjusts its behavior based on context, much like a person recalling a previous experience to guide their next move. This is crucial for real world tasks where conditions change constantly, such as handling objects of varying sizes or navigating cluttered spaces.

🆔 @hackernewsgithubprojects

📦 jprx/darwin-vm

Boot an iPhone Kernel on Your Computer

Boot a real iPhone or Mac right on your desktop. Darwin VM takes Apple's operating system and runs it inside a standard emulator, skipping the screen and apps entirely. In seconds, you land in a root shell with full control to debug the kernel and compile your own programs. This is huge because it opens up Apple's usually locked down architecture for deep research, letting you see exactly how their latest chips run the OS. If you have ever wanted to peek under the hood of an iPhone, this is your ticket. Give it a try.

🆔 @hackernewsgithubprojects

📦 raullenchai/twinspark

twinspark: The Self-Healing AI Cluster

Twinspark takes two modest desktop computers and turns them into a single, resilient engine capable of running massive language models locally. The project solves a critical pain point for local AI developers: the fragility of multi-node setups. By using a self-healing cluster architecture, the system automatically recovers from crashes or power outages without manual intervention, ensuring your AI assistant is always ready to work. It also fine-tunes the decoding process to squeeze out high performance from limited hardware, allowing you to handle complex coding tasks with speed and reliability.

🆔 @hackernewsgithubprojects

📦 zju-real/ttpo

TTPO: Making AI Math Better Without Human Labels

Boosting a language model's math skills without a single human-written answer is possible thanks to TTPO. This project lets AI improve its reasoning directly on test problems by learning from its own attempts. It solves a major issue where standard voting methods fail because most guesses are wrong. TTPO spots that disagreement is a useful signal. It rewards consistent answers and penalizes confident errors. This creates a self-improving cycle that works even without perfect data. The result is a model that gets significantly better at difficult math just by practicing on its own. No labels needed, just smart self-correction. That is a powerful way to build smarter AI.

🆔 @hackernewsgithubprojects

📦 zhangzaibin/future-robots

Future Robots: The Multi-Arm AI That Teaches Bots to Work Together

Future Robots is the robotics research code that finally makes multiple robot arms work as a coordinated team. It solves the problem of rigid, single-task training by allowing agents to decompose goals into small, reusable skills and compose them on the fly. Instead of treating every task as a unique, memorized routine, the system learns atomic actions that can be rearranged to handle new problems. This approach lets a group of robots collaborate on complex, multi-step tasks they were never explicitly trained to perform together.

🆔 @hackernewsgithubprojects

📦 ryanssenn/gemma4.c

The 700 Line LLM

One single C file runs a full language model on your CPU, outpacing standard optimized engines by more than double in speed. gemma4.c is an educational project that builds the entire Gemma 4 inference pipeline in just seven hundred lines of pure C, with zero external libraries. It strips away the heavy frameworks usually required, leaving only the raw math of tokenization, matrix multiplication, and attention. The goal is not to replace professional tools, but to make the internal mechanics of large language models completely transparent. By seeing every line of code that turns text into predictions, developers finally understand exactly how the magic happens under the hood.

📰 https://news.ycombinator.com/item?id=49468286

🆔 @hackernewsgithubprojects

📦 lilmgenius/paperthin

Paperthin: The Agent Skill That Deletes

Paperthin fixes the biggest problem with AI coding assistants: they never delete. Instead of just piling on more code and comments, these skills force your agent to clean up. One key skill rewrites a messy draft from scratch, stripping away the clutter to leave only what actually works. It turns old engineering wisdom into automatic habits that keep your projects lean and readable. It is like having a strict editor who refuses to add noise, ensuring your code stays clean instead of rotting into a mess. If you are tired of AI bloat, this is the tool that makes subtraction your default.

🆔 @hackernewsgithubprojects

📦 jalen-brunson/comfyui-minimax-h3-pdd-acc

Generate Video and Audio in Just Eight Steps

ComfyUI MiniMax H3 PDD Acc lets you generate synchronized video and audio in only eight computational steps, skipping the usual forty-plus iterations that slow down creation. It works by fusing a specialized acceleration layer directly into the model, allowing the system to predict large chunks of the final output at once rather than building it pixel by pixel. This setup removes the need for guidance scaling, making the process significantly faster while maintaining high quality. The tool handles both video and sound together, ensuring they stay perfectly in sync without extra post-processing.

🆔 @hackernewsgithubprojects

📦 seeray11/openflow

OpenFlow Visual AI Builder

OpenFlow is the visual canvas that lets you wire multiple AI agents together like a assembly line. Instead of writing complex code to manage different artificial intelligence tasks, you simply drag role cards onto a screen and connect them. The real magic happens when these agents run in parallel. A planner can think while an architect designs and a coder builds, all working on your actual files at the same time. It turns abstract AI concepts into a simple, visual workflow you can see and control. You get the power of teamwork without the headache of manual orchestration.

🆔 @hackernewsgithubprojects

📦 elliottdehn/open-jobs

Find Jobs with AI: open-jobs

Turn your dream job description into a living search engine with open jobs, a free tool that helps you hunt for roles without spending a dime. Instead of sifting through thousands of results, you simply describe the role you want to a coding agent, which then embeds that description to find the most similar live job postings from millions of listings. It builds a local, interactive page on your computer where you can rate jobs as you browse, training the system to refine its matches in real time.

🆔 @hackernewsgithubprojects

📦 agent4science-utokyo/task-coevolve

Task-CoEvolve: The 80% Faster AI Optimizer

Cut your AI testing costs by eighty percent without losing accuracy, all thanks to Task-CoEvolve. This University of Tokyo project solves the expensive problem of constantly re-evaluating every single task when improving an AI agent. Instead of grinding through the entire list, it smartly picks only the tricky problems where different AI versions actually disagree. By focusing solely on these useful moments, the system learns faster and stays accurate. It is a clever, efficient way to make your AI better while spending far less computing power.

🆔 @hackernewsgithubprojects

📦 whispem/asm.fm

asm.fm: Making Music with Pure Assembly

Make a song without a single audio library. That is exactly what asm.fm does. It turns raw CPU math into actual sound by calculating every note by hand and writing the numbers straight to a standard wave file. The project proves you can build a full synthesizer, complete with filters and effects, using only basic system calls. It strips away the complex software layers to show exactly how a speaker moves when fed with simple numbers. If you have ever wondered how computers create music from scratch, this project offers a fascinating look at the mechanics underneath.

🆔 @hackernewsgithubprojects

📦 james0248/visionary

Visionary: AI Agents That Dream in Browsers

Visionary is the research project that lets artificial agents learn physics by dreaming inside virtual worlds. Built on the Dreamer 4 architecture, it creates tiny, seven million parameter models that run directly in your web browser at thirty frames per second. The coolest part is how small these brains are yet still accurate enough to simulate complex robot movements like pushing objects or opening doors without falling apart. It’s a clever way to train general purpose agents without needing massive hardware or real world interaction. This repo shows that lightweight, high speed world models are finally ready for practical testing and creative experimentation.

🆔 @hackernewsgithubprojects

📦 omi-health/medical-stt-eval

Which AI Listens to Doctors Best?

Omi Health built a public benchmark that pits forty-two speech-to-text models against fifty-five realistic doctor and patient conversations to see who actually understands medical jargon. The surprise is not just who transcribes the fastest, but who gets the specific drug names right. They created a new score called Medical WER that isolates errors on clinical terms, revealing that a top general model might sound perfect to you but could miss a critical prescription detail. This project gives developers a clear, reproducible way to test if an AI scribe is safe for real healthcare use.

🆔 @hackernewsgithubprojects

📦 mickmumpitz/comfyui-sage-easyinstall

Sage Attention on Windows: The 2x Speed Hack

Getting Sage Attention running on Windows has historically been a total nightmare, but ComfyUI Sage Easy Install just turned that into a two-click job. If you are rendering video with ComfyUI, this tool automates the entire headache of finding the right compiler, matching your specific graphics card, and wiring up the code. It literally scans your system, picks the correct software pieces, and drops in a ready-to-go launcher. The best part is that it makes your video renders up to twice as fast without you having to touch a single command line. It even handles those annoying version mismatches that usually break things.

🆔 @hackernewsgithubprojects

📦 0xbakeer/qwen3.8-27b-fp8-on-a-single-dgx-spark

Qwen 27 Billion Runs Fast on One Tiny Device

The Qwen 27 Billion setup that makes a single compact device run fast is a fascinating engineering puzzle. The creator found a way to speed up this heavy AI model without changing the underlying weights, using clever tricks to predict future words. The most surprising result is that they squeezed fifty-eight tokens per second out of a small machine by guessing ahead, and only the eight-point format allows the newest prediction tools to work at all. This project proves that smart software tweaks can drastically improve speed on modest hardware. If you run AI locally, knowing the right settings can make a slow model feel instant without buying expensive equipment.

🆔 @hackernewsgithubprojects

📦 raysonmeng/agent-bridge

AgentBridge lets AI agents review each other

AgentBridge is the local bridge that finally lets two AI coding agents talk directly to each other without you copying and pasting text between terminals. It connects Claude Code and Codex so they can collaborate in real time, with one agent reviewing the other's work instantly. This solves the annoying problem where you have to manually relay messages between two separate AI sessions, acting like a human messenger. The coolest part is that these agents can negotiate who does what and even hand off tasks when one hits its usage limit, keeping your long projects moving overnight. No more manual ferrying of code snippets.

📰 https://news.ycombinator.com/item?id=49489198

🆔 @hackernewsgithubprojects

📦 studioplatforms/aistupidmeter-web

AI Stupid Meter: Is Your AI Model Secretly Getting Dumber?

AI Stupid Meter watches your AI models to catch when they quietly get worse. It uses a smart statistical method called CUSUM to spot sudden drops in performance that you might miss, flagging degradation before your code breaks or your users notice. Think of it as a smoke detector for your AI, telling you exactly when a model is failing so you can switch providers or tweak your prompts. This open source dashboard also helps you route requests to the best-performing model automatically, saving you money and headaches. It is a clever, practical tool for anyone serious about keeping their AI outputs reliable.

🆔 @hackernewsgithubprojects

📦 studioplatforms/aistupidmeter-api

The AI Stupid Meter API

The AI Stupid Meter is the backend system that finally catches AI models getting worse. It runs constant stress tests on major AI models to detect when they start making mistakes or acting inconsistently. This project uses statistical methods to spot slow degradation that simple reviews miss, alerting developers before a bad model breaks their production apps. It’s basically a quality control layer for artificial intelligence. Stop guessing which model is reliable; let this API track the data and flag the decline.

🆔 @hackernewsgithubprojects